A Portfolio Chatbot That Still Answers When Gemini Is Down

A chatbot on a portfolio is a nice extra until the model key is missing, the quota is gone, or the request times out. The widget should still answer “what have you shipped?” without a spinner that never ends.

On this site the chat route always returns a plain-text stream. Gemini fills it when it can. A local matcher fills it when Gemini returns nothing.

One response shape

The client does not care which brain wrote the tokens. Both paths send text/plain and Cache-Control: no-store. A response header tells me which one ran:

if (geminiStream) {
  return new Response(geminiStream, {
    headers: {
      "Content-Type": "text/plain; charset=utf-8",
      "X-River-Chat": "gemini",
    },
  })
}
 
return new Response(localStream, {
  headers: {
    "Content-Type": "text/plain; charset=utf-8",
    "X-River-Chat": "local",
  },
})

streamGeminiReply returns null when there is no GOOGLE_GENERATIVE_AI_API_KEY, the fetch throws, the status is not ok, or the body is empty. Null means fall through. It does not mean show an error toast.

What Gemini is allowed to say

The system prompt is the portfolio, not the open web: about, experience, projects, awards, and the blog list. If the answer is not in that text, the model is told to say it doesn’t know and link /contact. Hiring questions go to /hire.

History is capped at the last 12 turns. Output is capped. The request aborts after 20 seconds. A hung model should not hold the function open.

The local matcher

Without a model I still need an answer in English or Vietnamese. The fallback scores the latest user message against phrase lists: greeting, projects, stack, experience, hire, contact, resume, blog, now, social.

JavaScript \b breaks on Vietnamese letters, so the patterns use a Unicode lookaround instead of a word boundary. The highest-weight hit wins. No hit gets a short fallback that points at real pages.

The local reply is then streamed word by word with a tiny delay, so the UI still feels like a reply instead of a pasted paragraph.

Guard the route first

Neither backend runs until the request is boring and small:

• 20 messages per IP per minute, then 429
• Zod: at most 24 messages, locale en or vi
• Latest user text sliced to 2000 characters
• Empty text is 400

The local matcher is not a jailbreak. It only knows this site. Gemini is told the same boundary. A visitor asking for a general coding tutorial gets a refusal and two example questions about the portfolio.

When to upgrade

This is enough for a personal site. It is the wrong shape for a support bot that must remember tickets or call tools. If the local answers start feeling thin, add facts to the profile text both backends already read — don’t add a third model.

Related on this site

• Rate limiting Next.js routes — the 20/min chat ceiling
• Counting blog views once — another public counter with a slug allowlist
• Visitor stats without a disk — Redis vs file for longer-lived numbers
• Rebuilding this portfolio — how the chat widget landed here